home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.parallel
- Path: sparky!uunet!europa.asd.contel.com!gatech!hubcap!fpst
- From: rvdg@cs.utexas.edu (Robert van de Geijn)
- Subject: Broadcasting on Hypercubes
- Message-ID: <l6bna3INNjjh@grit.cs.utexas.edu>
- Keywords: iPSC Hypercubes Broadcast
- Sender: fpst@hubcap.clemson.edu (Steve Stevenson)
- Organization: CS Dept, University of Texas at Austin
- Date: 16 Jul 1992 15:34:11 -0500
- Approved: parallel@hubcap.clemson.edu
- Lines: 75
-
- I recently developed a broadcast routine for the iPSC/860
- that is considerable faster than the library call to
- csend with destination -1, at least for very large vectors.
-
- While I am still working on the report, and I am not willing
- to disclose the algorithm at this time, I am willing to
- make the object code available to whoever needs it in a
- hurry.
-
- A time complexity shows that the algorithm is faster that
- a minimum spanning tree broadcast (which is used by csend)
- by a factor (p log p) / (2 (p-1) ) (asymptotically).
-
- Here are some timings on the ORNL iPSC/860:
- (n= # of bytes, p= # of nodes, times in seconds)
-
- p=64
- n new csend-1
- = === =======
- 0 0.00020 0.00022
- 1 0.00021 0.00022
- 10 0.00022 0.00024
- 100 0.00041 0.00041
- 500 0.00210 0.00188
- 1000 0.00263 0.00295
- 5000 0.00645 0.01163
- 10000 0.01021 0.02250
- 50000 0.04000 0.10796
- 100000 0.07391 0.21513
-
- p=32
- 0 0.00017 0.00018
- 1 0.00018 0.00018
- 10 0.00019 0.00020
- 100 0.00034 0.00034
- 500 0.00163 0.00155
- 1000 0.00239 0.00245
- 5000 0.00622 0.00960
- 10000 0.00926 0.01854
- 50000 0.03779 0.08996
- 100000 0.07202 0.17925
-
- The object file can be obtained via anonymous ftp
- from cs.utexas.edu, directory pub/rvdg/gcast.
- In this directory, there is a README file that
- explains the calling sequence.
-
- ===========================
- ftp cs.utexas.edu
- Name: anonymous
- Password: anything
- cd pub/rvdg/gcast
- bin
- get rvdg_gcast.o
- get README
- ===========================
-
- you may want to look at some of the postscript files
- in pub/rvdg as well.
-
- Those who send a message to rvdg@cs.utexas.edu will
- be kept up to date on further developments wrt this
- and other communication algorithms.
-
- We are working on a broadcast for the Intel Touchstone
- Delta as well, based on the same principles.
-
- PLEASE, don't send messages asking for further details.
-
- Robert van de Geijn
- Assistant Professor
- The University of Texas at Austin
- rvdg@cs.utexas.edu
-
-
-